Medical image screening that explains itself using past cases

Most medical image screening models treat each case in isolation and rely on hard-to-interpret saliency maps generated after the fact. EviScreen retrieves relevant regions from a dual knowledge bank of historical cases, feeding that evidence directly into its prediction process rather than appending an explanation afterward. The result is both better performance — notably higher specificity at clinical-level recall on real-world screening benchmarks — and built-in localization interpretability via contrastive retrieval abnormality maps. Code is publicly available on GitHub, making this accessible to medical AI practitioners.